Creating a predictive model for mastitis with dairy digital ag data

Your data comes from CAST, which is creating an ecosystem of networked technologies and techniques for modern farms. What kind of data did they give you?

We're using data gathered by CAST from their main farm, the Cornell Dairy Research and Education Center (CDREC). There’s a lot of different kinds. We have herd health data from Dairy Comp, a dairy herd management software. We also have diagnostic data from the Cornell Animal Health Diagnostic Center for CDREC. Then, we have milk parlor data from Delpro, which is a dairy and milking management platform. Delpro gives us all kinds of measurements of milk output and composition. And we have geospatial ear-tag data, which shows the location of each cow within the barn at a millisecond level.

CDREC is our first farm, and we're excited to get a testable prototype through the data CAST is providing.

Can you describe your research and your process for creating the prediction model?

We’re starting with prediction of mastitis episodes. So, we’re looking at when a cow gets mastitis and the factors that show us she still has it, such as reduced milk yield or a larger percentage of low milk flow. Then, we use these models as a baseline to indicate the top contributing factors. 

We then compare the outputs from the model, and what the model considers important, such as traditional biological factors like the sum of historical mastitis cases for a particular cow. Whenever the model has stronger signals than we expect, or when the model is giving us unexpected signals, we can combine that with the biological signals to give us insights. 

For example, we might infer that lining slips are causing an increased risk of mastitis on a farm. This happens  when the milking machine loses its tight seal, allowing a cow’s teats to be exposed to air during the milking process. We can recommend updating automated milking machine settings, such as the vacuum power, to prevent this. 

Or, if we can identify areas of the barn where cows are at a higher risk, we can walk through the barn and try to associate that increased risk with an influential factor and a fix. For example, we might notice that there’s water pooling in a specific area of the barn and that it’s developing a persistent pathogen or algae that is increasing the risk of mastitis. 

Overall, we want to develop a locally adapted, flexible system where we can tailor a model to any farm and use whatever data they have available to give them predictive power and actionable insights.

It seems like the model you’re creating should increase herd productivity and profitability if farmers use it to deal with mastitis in a more timely manner. 

Yes, that’s what we are aiming for. We want to provide actionable insights from the data the farmer has and help them find things like malfunctioning machines or identify if a specific section of the barn needs new bedding, for example. If our model can give them insights on ways to increase the efficacy of targeted interventions or treatments based on the cow’s mastitis risk, that should help raise productivity and profitability—and increase cow welfare. The models are limited to the data the farmer can provide, but the goal is to take whatever data the farm has and make it work for them. 

Looking forward, what are the next steps?

We have two big next steps! Right now, we are working on two papers. One is about mastitis prediction from parlor, weather, herd health and diagnostic data. The other is about mastitis prediction based on geospatial data.  

Jackie Swift is the communications specialist for the Cornell CALS Department of Animal Science and the communications manager for CAST.